EFTA00307750.pdf

NEWS IN FOCUS CANINE PSYCHIATRY Dogs provide PHYSICS Debate over meaning CLIMATE Developingnations CalSERVATMI Songbird killing genetic clues to human of Stephen Hawking's latest struggle to keep carbon for restaurants becomes a disorders µe paper µ41 accounts p.451 hot issue in Cyprus µS2 Go. a complex game popular in Asia. has I ustrated the efforts of artificial-intelligence researchers fordecades. ARTIFICIAL INTELLIGENCE Google masters Go Deep-learning software excels at complex ancient board game. BY ELIZABETH GIBNEY reveals in research published in Nature on famously beat grandmaster Garry Kasparov in 27 January'. It also defeated its silicon-based 1997, was explicitly programmed to win at the computer has beaten a human rivals, winning 99.8% of games against the game. But AlphaGo was not preprogrammed A professional for the first time at Go — an ancient board game that has long been viewed as one of the greatest challenges current best programs. The program has yet to play the Go equivalent of a world chant- pion, but a match against South Korean pro- to play Go: rather, it learned using a general- purpose algorithm that allowed it to interpret the game's patterns, in a similar way to how a for artificial intelligence (Al). fessional Lee Sedol, considered by many to be DeepMind program learned to play 49 different The best human players of chess, draughts the world's strongest player, is scheduled for arcadegames'. and backgammon have all been outplayed by March. "We're pretty confident," says Deep- This means that similar techniques could be computers. But a hefty handicap was needed Mind co-founder Demis Hassabis. applied to other AI domains that require recog- for computers to win at Go. Now Google's "This is a really big result, it's huge," says Kenn nitionofcomplex patterns, long-term planning London-based Al company, DeepMind, claims Coulom, a programmer in Lille, France, who and decision-making, says Hassabis. "A lot of that its machine has mastered the game. designed a commercial Go program called the things we're trying to do in the world come DeepMind's program AlphaGo beat Fan Crazy Stone.1-lehad thought computer mastery under that rubric.' Examples are using medical Hui, the European Go champion, five times of the game was a decade away. images to make diagnoses or treatment plans, out of five in tournament conditions, the firm The IBM chess computer Deep Blue, which and improvingclimate-change models. 29 JANUARY 2016 I VOL 529 I NATURE I 445 2016 Macmillan Publishers Lmited. All rights reserved EFTA00307750  NEWS IN FOCUS ► In China, Japan and South Korea. Go is other programmes categorize images front have developed a conservative (rather than hugely popular and is even played by celebrity pixels (see Nature 505, 146-148; 2014). Then aggressive) style, adds TobyManning, a lifelong professionals. But the game has long interested it played against itself across 50 computers, Go player who refereed the match. Al researchers because of its complexity. The improving with each iteration, a technique Google's rival firm Facebook has also been rules are relatively simple: the goal is to gain known as reinforcement learning. working on software that uses machine learn- the most territory by placing and capturing The software was already competitive with ing to play Go. Its program, called darkforest, black and white stones on a 19 x 19 grid. But the leading commercial Go programs, which is still behind commercial state-of-the-art Go the average 150-move game contains more select the best move by scanning a sample of Al systems, according to a November preprint'. possible board configurations — l0p0 — than simulated future games. DeepMind then com- Hassabis says that many challenges remain there are atoms in the Universe, so it can't be bined this search approach with the ability to in DeepMind's goal ofdevelopinga generalized solved by algorithms that search exhaustively pick moves and interpret Go boards — giving Al system. In particular, its programs cannot for the best move. AlphaGo a better idea yet usefully transfer their learning about one "Deep learning ofwhichstrategies are system— such as Go— to new tasks a feat that AISTRACT STRATEGY is killing every likely to be success- humans perform seamlessly. "We've no idea Chess is less complex than Go, but it still has too problem in AL" M. The technique is how to do that. Not yet," Hassabis says. many possible configurations to solve by brute "phenomenal", says Go players will be keen to use the software force alone. Instead, programs cut down their Jonathan Schaeffer. a computer scientist at the to improve their game, says Manning, although searches by looking a few turns ahead and judg- University of Alberta in Edmonton, Canada, Hassabis says that DeepMind has yet to decide ing which player would have the upper hand. In whose software Chinook solved' draughts in whether it will make a commercial version. Go, recognizing winning and losing positions 2007. Rather than follow the trend of the past AlphaGo hasn't killed the joy of the game, is much harder: stones have equal values and 30 years oftrying to crack games using comput- Manning adds. Strap lines boasting that Go is can have subtle impacts far across the board. ing power, DeepMind has reverted to mimick- a game that computers can't win will have to To interpret Go boards and to learn the best ing human-like knowledge, albeit by training, be changed, he says. "But just because some possible moves, the AlphaGo program applied rather than by beingprogrammed,he says. The software has got to a strength that I can only deep learning in neural networks — brain- feat also shows the power ofdeep learning, which dream of, it's not going to stop me playing:' • inspired programs in which connections is going from success to success, says Coulom. SEE EDITORIAL P.437 between layers of simulated neurons are "Deeplearning iskillingevery probleminAl" strengthened through examples and experi- AlphaGo plays in a human way, says pan. 1. Silver. D. et aL Nature 529.484-489(201e. ence. It first studied 30 million positions from "If no one told me. maybe I would think the 2. Mnih, V. eta,. Nature 518.529-533 (2015). 3. Schaeffer. J. eta/ Seknce 317,1518-1522 (2007) expert games, gleaning abstract information player was a little strange, but a very strong 4. Tian, Y. & Zhu. Y. Preprint at arXi http://arxiv.org/ on the state ofplay from board data, much as player, a real person." The program seems to pdf/1511.06410.pdf (2015). GENOMICS Dog DNA probed for clues to human psychiatric ills Project will compare gene data to owners' assessments of how their companions behave. BY HEIDI LEDFORD determine why some dogs are more prone to the genetic links to human psychiatric disorders disorder than others. Her owner, Marjie Alonso by analysing DNA samples from thousands ddie plays hard for an II -year-old of Somerville, Massachusetts, has enrolled her of people. Those efforts have in recent years A greater Swiss mountain dog — she will occasionally ignore her advanced years to hurl her 37-kilogram body at an in a project called Darwin's Dogs, which aims to compare information about the behaviour of thousands of dogs against the animals DNA met with some success in schizophrenia and depression. But for some conditions, includ- ing OCD, not a single robust genetic link has unwitting house guest in greeting. But she profiles. The hope is that genetic links will been sifted from the background noise of carries a mysterious burden: when she was emerge to conditions such as canine compul- normal genetic variation. 18 months old, she started licking her front sive disorder and canine cognitive dysfunction Human studies are difficult in part because legs aggressively enough to wear off patches — a dog analogue of dementia and possibly the species is so genetically diverse, says of fur and draw blood. Alzheimer's disease. The project organizers have Wynne. Dogs, however, are more genetically Addie has canine compulsive disorder — a enrolled 3,000 dogs so far, but hope to gather homogeneous. Selected over thousands of condition that is thought to be similar to data from at least 5,000, and they expect tobegin years for particular characteristics, they dis- human obsessive-compulsive disorder analysingDNA samples in March. play less genetic variation than do humans. (OCD). Canine compulsive disorder can cause "It's very exciting, and in many ways it's Pure-bred dogs, in particular, have been ren- dogs to chase their tails for hours on end, or way overdue says Clive Wynne, who studies dered highly genetically consistent to achieve a to suck on a toy or body part so compulsively canine behaviour at Arizona State University homogenous appearance and behaviour. that it interferes with their eating or sleeping. in Tempe. Dogs also live side-by-side with humans, Addie may soon help researchers to Researchers have long struggled to find which some think can make them a better 446 I NATURE I VOL 529 1 28 JANUARY 2016 ft 2016 Macmillan Publishers Ignited. All rights reserved EFTA00307751